AI transparency does not end with a single disclosure

A generic footer creates one visible location, but it does not determine whether the actor is a provider or deployer, whether the object is an interaction or generated output, or who controls first exposure. Article 50 separates transparency by actor, system or output, audience, timing, and medium rather than treating it as one UI component.[1]

The non-obvious point is that human-facing disclosure and machine-readable marking are not stronger and weaker versions of the same device. Provider-side marking under Article 50(2) enables software to read and detect generation or manipulation. Deployer-side disclosure under Article 50(4) makes certain deepfakes and other content intelligible to people. Evidence of one implementation does not replace the decision for the other.[1] [3]

This article treats the release question not as whether a label exists, but whether every applicable role-and-object combination has the correct surface and evidence. Duties can apply cumulatively to the same system or group of actors, so the branching decision must precede the choice of badge.[3]

Article 50(1) requires the provider of an AI system intended to interact directly with natural persons to inform them that they are interacting with AI. The law includes bounded qualifications, including when that fact is obvious in the circumstances, but when the duty applies Article 50(5) makes delivery no later than the first interaction or exposure part of the design.[1] [3]

Article 50(2) requires providers of systems that generate synthetic audio, image, video, or text to mark outputs in a machine-readable format and make them detectable as artificially generated or manipulated. Article 50(3), by contrast, requires deployers of emotion-recognition and biometric-categorisation systems to inform exposed natural persons.[1]

Article 50(4) separates two further deployer duties: disclosure of deepfakes and disclosure of certain AI-generated or manipulated text published to inform the public on matters of public interest. Article 50(5) makes clear and distinguishable information, timing, and accessibility horizontal requirements. Paragraphs (1)–(4) are not mutually exclusive menu choices.[1] [3]

Machine-readable marking is not synonymous with a badge people see on every output. The non-binding Commission guidelines describe software-extractable information and perceptible disclosure as complementary and do not derive a universal visible-label duty for every synthetic output from Article 50(2). Marking also involves effectiveness, interoperability, robustness, and reliability so far as technically feasible.[1] [3]

Conversely, a deployer publishing a deepfake cannot treat provider metadata alone as the human-facing disclosure. The wording, first exposure, and placement in the screen, player, or text must be designed so the actual audience can distinguish it. Burying the information in terms, a manual, or a footer is not the same as disclosure at first exposure.[1] [3]

This article uses a two-layer handoff: receive the marking method, detection method, supported format, and version from the provider, then have the deployer check whether the mark survives distribution and place its human-facing disclosure. The final Code gives signatories concrete marking, detectability, visible or audible disclosure, and internal-process measures without merging the two responsibilities.[3] [4]

Article 50(1) includes bounded qualifications for interactions that are obvious in the circumstances and certain legally authorised criminal-law uses. This article's release check does not treat a product owner's assumption that users will surely understand as sufficient. It checks the intended audience, contact surface, and opening of the interaction in the release candidate and retains the factual basis for obviousness.[1] [3]

Article 50(2) expressly addresses standard-editing assistance, outputs that do not substantially alter input data or its semantics, and certain legally authorised law-enforcement uses. The guidelines' treatment of closed products and ephemeral real-time generation is non-binding interpretation of proportionality and technical feasibility, not a set of new blanket statutory exceptions.[1] [3]

For deepfakes in evidently artistic, creative, satirical, fictional, or analogous works, Article 50(4) attenuates the manner of disclosure but still contemplates an appropriate disclosure that does not hamper display or enjoyment. Public-interest text has an exception involving human review or editorial control plus editorial responsibility for publication. A superficial review should not be treated as automatic proof of the exception.[1] [3]

• Applicability: Fix the actor, system or output category, audience, and publication or exposure facts first
• Express exception: Retain the paragraph wording, relevant facts, scope, and decision owner in one record
• Attenuated disclosure: Do not confuse it with no duty; test whether the audience is informed without hampering enjoyment of the work
• Guideline interpretation: Keep it separate from statutory exceptions and retain the version and non-binding rationale

Under Article 113 of the AI Act, Article 50 has applied since 2 August 2026, and the 2026 amendment left that date and the core structure of the duties in place. As of 26 September 2026, 2 December (the Article 111(4) transition) is the only remaining Article 50 date; reading it as a general grace period would delay interaction notices and deployer disclosures that already apply.[1] [2]

The transition added to Article 111(4) by Regulation (EU) 2026/1744 is narrow. Providers of synthetic-content-generating systems placed on the market before 2 August 2026 have until 2 December 2026 to comply with Article 50(2). It is neither a general postponement of Article 50 nor a blanket grace period for new systems.[2]

Where one product both interacts and generates content, separate the Article 50(2) marking that may qualify for the transition until 2 December from the Article 50(1) interaction notice that has applied since 2 August. The Commission guidelines' treatment of outputs generated before 2 August as not requiring retroactive action is interpretation, not operative language carrying the same legal weight as Article 111(4).[2] [3]

The Commission guidelines published on 20 July 2026 are influential implementation material on roles, cumulative application, placement, and technical feasibility, but they are non-binding. The library page says they were adopted, while C(2026) 5054 says the content was approved and formal adoption would follow once all language versions were available. Because no separate later adoption instrument had been identified as of 26 September 2026, this article uses the narrower description: non-binding guidelines whose content the Commission approved.[3]

The final Code dated 10 June 2026 is voluntary and provides a signatory route for Article 50(2), (4), and (5). It does not cover the interaction notice in paragraph (1) or the emotion and biometric notice in paragraph (3), and it mixes signatory commitments with encouraged or optional measures. Following it cannot be expanded automatically to every Article 50 duty or treated as the only implementation method.[4]

The Commission Opinion and AI Board assessment found the Code adequate, but adherence is not conclusive evidence of compliance. The Board retained concerns about interoperability, short text, and downstream label retention and noted the need for reassessment where measures prove insufficient, impractical, or disproportionate. Amended Article 50(7) likewise preserves Board-informed assessment and the possibility of common rules if the Code is inadequate.[2] [5]

Catalysphere recommends deriving a disclosure placement record from the role-output matrix rather than copying the statutory text into a form. This is neither a prescribed legal document nor a certificate of conformity. It is an internal control that makes it possible to recheck which version was used, who assessed each duty, and where the disclosure or marking was implemented.[1] [3] [4] [5]

The core of the record is the decision field, not the disclosure wording. Link the system and release ID, provider or deployer role, Article 50 paragraph, system or output category, audience, first interaction or exposure, human-facing medium, machine-readable method, exception or attenuated-treatment rationale, accessibility, technical limit, evidence, owner, and review date. Do not release with applicable, not applicable, exception, or attenuated treatment left blank.

At the release gate, inspect the human-facing disclosure in the rendered candidate, extract and detect the mark from actual outputs when applicable, and test the assumption that it survives downstream transformation. Re-review when the legal source, Code, format, model, or distribution surface changes, and name the stop or rollback owner for missing disclosure or failed marking. The gate does not automate legal judgment; it stops unresolved branches from disappearing into release.

• Current source: Fix the versions and check dates for the consolidated law, amendment, guidelines, Code, Opinion, and Board assessment
• Role and duty: Decide the provider or deployer role, Article 50 paragraph, and system or output category separately
• Human-facing surface: Render and inspect the audience, wording, channel, placement, first interaction or exposure, and accessibility
• Machine-readable surface: When applicable, inspect the format, embedding point, extraction, detectability, and interoperability target
• Exception path: Distinguish an express exception, attenuated disclosure, and technical-feasibility interpretation; retain facts, authority, and approver
• Evidence owner: Link screenshots, sample outputs, inspection results, downstream tests, and the review date to the release ID
• Change and rollback: Name triggers for model, format, UI, distribution, and legal-source changes, plus owners for stop, repair, and republication